# Transcription

> Choose the language and model Samurai Scribe transcribes with, try two models side by side, and pick whether it runs on your graphics card or processor.

- Page: https://samuraiscribe.com/docs/transcription/
- Checked with Samurai Scribe 0.65.0

Samurai Scribe turns speech into text on your computer, for meetings, live captions and dictation. Everything about it is set in one place, the **Transcription** card in **Settings → AI Models**: the language spoken, the model that writes the text, the captions you see while someone speaks, and whether the work runs on your graphics card (the GPU) or your processor (the CPU). Running on the GPU is free on every plan.

## The Transcription card

**Settings → AI Models** opens on the **Transcription** card. It has two columns, **Meetings** and **Dictation**, and reads from top to bottom in the order the choices depend on each other:

1. **Language**: what is spoken in meetings, and what you dictate in.
2. **Translate to English**, which only Whisper models can do.
3. **Mode**: **As you speak** or **When you finish**.
4. **Model**: the model that writes the text. Click it to open **Transcription models**, the list to download and choose from.
5. **Compare models…**, to try two models on the same audio.
6. **Advanced**, which holds **Run transcription on**.
7. **Captions**, in a block of their own, while either column is set to **As you speak**.

Each row's ⓘ says what it does. **Link** and **Unlink**, at the end of the **Model** row, decide whether meetings and dictation share one model or each have their own.

Below the Transcription card, the **Dictionary** card lists the models that listen for your own terms in the audio, with **Open Dictionary**. See [the Dictionary](https://samuraiscribe.com/docs/dictionary/).

## Two languages: meetings and dictation

Speech-to-text has two languages, set side by side on the **Language** row:

- **Meeting language** is used for recordings, live captions and uploaded audio. **Auto-detect** usually works. Name the language when your meetings open with small talk in another one.
- **Dictation language** is used for both kinds of dictation.

The two are never linked, so you can meet in French and dictate in English with the same model.

Every language is listed, whatever model you have. If the model can't transcribe the language you chose, a line under the model says so, such as "Parakeet v2 (English) only transcribes English", with a button to use or download one that can. Nothing changes until you click it.

Until you do:

- a meeting's **Run** says why it can't start;
- a dictation says so on screen, and names **Settings → AI Models**, instead of typing words in the wrong language or quietly switching to **Auto-detect**.

## Which models are free

Choose a model in **Transcription models**, which opens when you click the model in the **Meetings** or **Dictation** column. **NVIDIA Parakeet** models are the most accurate, and can't translate. **Whisper** models cover 99 languages and can translate into English.

### Free and Pro models

| Model | Plan |
|---|---|
| Parakeet v2 (English), Parakeet v3 (Multilingual) | Free |
| Whisper Tiny, Base and Small | Free |
| The captions models, English and Multilingual | Free |
| Parakeet Ultra (Multilingual) | Pro |
| Whisper Medium, Large V3 Turbo and Large V3 | Pro |

On first run the app picks Parakeet v2 if you speak English, Parakeet v3 for its other languages, and Whisper Small for any other language.

On the free plan, a Pro model can still be downloaded and tried. Its row offers **Download to try**, and once it is downloaded, **Try it**, which opens [Compare models](#compare-models) with that model beside your meetings model. Using it for meetings or dictation needs Pro.

### When a download fails

A model whose download failed says why under its name, with **Try again**.

### Parakeet's new engine

Since version 0.65, Parakeet can use your graphics card. A Parakeet model downloaded before 0.65 needs one more file for this. The first time you open the new version, **Update Parakeet to its new engine** lists the models that need it:

- **Update** downloads the files in the background, one after another. Each model keeps working as it is until its file is ready.
- **Not now** closes it, and it doesn't ask again. Each model's row in **Transcription models** still offers **Get the GPU file**, with its size.

Until it is updated, that model runs on the CPU only.

## Captions

While a column is set to **As you speak**, the **Captions** block shows what the captions model puts on screen while someone speaks. Each caption is replaced by what your transcription model writes once the turn is over. Nobody chooses the captions model; the language does. English uses the English captions model once it is downloaded, and any other language the multilingual one. While English captions use the multilingual model, a line under them offers the English one.

**Captions memory**, inside the block, sets when the captions models are loaded into memory and released.

## Run transcription on

Open **Settings → AI Models**, then **Advanced** in the **Transcription** card. **Run transcription on** has three choices:

- **Automatic**, the default. The app checks each transcription model you use on both the CPU and the GPU, and uses whichever was faster for that model.
- **GPU** always uses the graphics card, for every model that can run there.
- **CPU only** always uses the processor. Choose it if transcription fails, or if the computer misbehaves while it runs.

Most of the time the GPU is faster, but not always: some graphics cards, such as the ones built into the processor, are slower than the processor itself. That is why Automatic checks.

Until a model has been checked, Automatic runs it on the GPU, except on an Intel Mac, or a Windows PC whose only graphics are built into the processor, where it starts on the CPU.

The words in the transcript are the same on either one; only the time it takes changes. The setting applies everywhere the app transcribes: meetings, the text that replaces live captions, and dictation.

While **Advanced** is open with **Automatic** chosen, a line for each model in use says where it runs and when it was checked, or that it has not been checked yet. **Check again** times them again, for example after a driver update. The check runs only while nothing is recording or processing, and once more after each app update. If you choose **GPU** or **CPU only**, **Advanced** stays open, so the choice is always in sight.

If no model you use can run on the GPU, **GPU** can't be chosen, and says "Your models run on the CPU on this computer".

### On Windows

Whisper and Parakeet both run on the GPU on Windows, or on the CPU. On a PC with both built-in graphics and a separate graphics card, the app uses the separate card.

On an older processor that Whisper models can't run on, Parakeet still works, on the CPU.

## Compare models

**Compare models** runs the same audio through two models, or one model on the CPU and on the GPU, one after the other on your computer. Open it with **Compare models…** in the **Transcription** card, or **Compare models** at the top of **Transcription models**.

1. Under **Audio**, choose what to run on:
   - **Sample meeting**, a short real meeting that comes with the app, with a transcript of what was said;
   - one of your meetings;
   - **Read a passage aloud…**, which records you reading a passage shown on screen.
2. Choose **Meeting** or **Dictation**, to test the model the way a meeting or a dictation uses it.
3. In each column, pick a model, and **CPU** or **GPU** when it can run on both.
4. Press **Run**.

Each column shows how long its model took, and on the sample or a passage you read, how many words it got wrong. On one of your meetings there is nothing to score against, so the columns are compared with each other. A line under both says which was faster and which got fewer words wrong, and **Where they differ** lists the places their text differs. Nothing is saved.

When the two columns are the same model on the CPU and on the GPU, and one is clearly faster, **Use the faster one** keeps that device for the model and sets **Run transcription on** to **Automatic**.

On the free plan, a Pro model can be compared on the sample as often as you like. On your own meetings or a passage, it gets a shorter stretch of audio and a few comparisons each time you open the app.